Eiraj Saqib försvarade framgångsrikt sin doktorsavhandling

Opponent på avhandlingen var professor Holger Fröning, Heidelberg University, Tyskland. Tillsammans med betygskommittén bestående av professor Slawomir Nowaczyk, Högskolan i Halmstad, professor Giandomenico Licciardo, University of Salerno samt docent Qing He, Mittuniversitetet, granskades och godkändes arbetet noggrant. 

Avhandlingen handleddes av professor Mattias O'Nils och Dr. Irida Shallari vid Mittuniversitetet.

Abstract

The proliferation of the Internet of Things (IoT) necessitates deploying Deep Learning (DL) models, specifically Convolu[1]tional Neural Networks (CNNs), on resource constrained edge devices. However, the high computational and memory de[1]mands of CNNs often exceed the capabilities of IoT nodes, while traditional cloud offloading suffers from latency and bandwidth limitations. This thesis proposes a comprehensive framework for Split Computing, enabling efficient distributed inference by partitioning CNNs between IoT nodes and edge servers.

The core contribution is a bottleneck-aware feature com[1]pression mechanism designed to minimize data traffic at the partition point. The research demonstrates that combining partitioning with extreme quantization (down to 1-bit) and compression reduces data transmission by over 99% with min[1]imal accuracy loss. This approach is augmented by a novel hybrid structured pruning criterion, utilizing L2-norm magni[1]tude and entropy, which selectively removes non informative channels to achieve significant speed-ups and energy savings compared to baseline execution modes.

To address quantization induced accuracy degradation, the thesis introduces Time-dependent Clustering Loss (TCL), a regularization technique that clusters activations during train[1]ing to ensure robustness against extreme quantization errors.

Furthermore, the complex selection of partition points, compres[1]sion ratios, and quantization levels is automated via CO-NAS (Compression Optimization, and Neural Architecture Search), a differentiable architecture search framework that efficiently discovers Pareto-optimal configurations.

Validated on diverse hardware platforms (e.g., Raspberry Pi, NVIDIA Jetson) and standard datasets (CIFAR-100, Tiny[1]ImageNet), these methodologies establish a robust pathway for Edge Intelligence. By unifying partitioning, quantization, pruning, and automated search, this work provides a scalable solution for deploying high performance vision models in resource constrained IoT environments.

Läs hela avhandlingen här